Nae-Chyun Chen

Supplementary Table S1. Training set sensitivity and specificity across confounders

Supplementary Figure S2. a) Model scores increase with overall stage (left panel), T stage (middle panel) and N stage (right panel). Statistical significance is determined using a one-sided Mann-Whitney U-test. b) Overall sensitivity and specificity across different subgroups at 90% specificity in the training set. Dashed lines indicate the overall sensitivity and specificity. Confidence intervals are derived using the Pearson-Clopper method. c) The model’s sensitivity remains robust across tumor locations within the colon and rectum in the training set. Confidence intervals are calculated using the Pearson-Clopper method d) The model’s sensitivity remains robust across tumor types. Confidence intervals are calculated using the Pearson-Clopper method

Supplementary Figure S1) To assess the robustness and reproducibility of the CRC prediction models, we evaluated their performance on samples processed under varying conditions. This figure shows the performance of the final CRC model, after retraining on the full training dataset, across the robustness and reproducibility studies. Specifically, we analyzed the precision of model scores using a set of precision study samples processed by two different operators, instruments, and, when available, reagent lots. Two separate precision studies were performed: One that started from separate plasma aliquots and tested variability derived from extraction. For this study, once RNA was extracted, it was included in the same batch for processing through the rest of the assay (left panel). Across the replicates, the scores remained consistent, with no instances where changes in scores were significant enough to cross the cutoff threshold into the opposite category. The second study used pools of RNA that were diluted 2-fold to generate more input material and then distributed in replicates across different batches starting from library preparation and through the rest of the assay (middle panel). These experiments ensured that normal laboratory variability did not significantly impact the model’s predictive outputs. Finally, we compared the scores of 40 samples where 2 separate plasma aliquots were processed at different times through the assay ( Right panel). The high correlation of model scores across these conditions demonstrates the model’s robustness to technical variability and underscores its reliability for real-world applications

Supplementary Table S2. Validation set sensitivity and specificity across confounders

Purpose: Colorectal cancer is the second leading cause of cancer-related deaths worldwide, and early detection significantly improves treatment outcomes, but existing blood-based tests often have limited sensitivity in early-stage disease. We developed a blood-based test combining orphan noncoding RNAs (oncRNA), a group of small cell-free RNAs, with generative artificial intelligence to detect colorectal cancer. Experimental Design: We leveraged a cohort of 613 colorectal cancer cases and controls to train a model that demonstrated both high clinical performance and minimal technical variability in robustness testing. We further validated our model in an independent, single-source cohort of 192 colorectal cancer cases and controls. Model performance was assessed by sensitivity, specificity, and area under the ROC curve, with attention to early-stage detection. Results: In our independent validation set, we achieved an overall sensitivity of 89% at 90% specificity, with an 80% sensitivity for stage I—an important milestone, as early-stage colorectal cancer detection remains a challenge for other blood-based technologies. Performance was consistent across demographic subgroups. Conclusions: Our oncRNA–based blood test, powered by artificial intelligence, offers strong performance for early colorectal cancer detection, including in stage I disease for which existing blood-based assays are limited. These findings support further development toward a minimally invasive colorectal cancer screening tool.

Cell-free RNA (cfRNA) profiling has emerged as a powerful tool for non-invasive disease detection, but its application is limited by data sparsity and complexity, especially in settings with constrained sample availability. We introduce Exai-1, a multi-modal, transformer-based generative foundation model that integrates RNA sequence embeddings with cfRNA abundance data to capture biologically meaningful representations of circulating RNAs. By leveraging both sequence and expression modalities, Exai-1 captures a biologically meaningful latent structure of cfRNA profiles. Pre-trained on over 306 billion tokens from 8,339 samples, Exai-1 enhances signal fidelity, reduces technical noise, and improves disease detection by generating synthetic cfRNA profiles. We show that self-attention and variational inference are particularly important for preservation of biological signals and contextual relationships. Additionally, Exai-1 facilitates cross-biofluid translation and assay compatibility through disentangling biological signals from confounders. By uniting sequence-informed embeddings with cfRNA expression patterns, Exai-1 establishes a transfer learning foundation for liquid biopsy, offering a scalable and adaptable framework for next-generation cfRNA-based diagnostics.

Liquid biopsies have the potential to revolutionize cancer care through non-invasive early detection of tumors, when the disease can be more effectively managed and cured. Developing a robust liquid biopsy test requires collecting high-dimensional data from a large number of blood samples across heterogeneous groups of patients. We propose that the generative capability of variational auto-encoders enables learning a robust and generalizable signature of blood-based biomarkers that capture true biological signals while removing spurious confounders (e.g., library size, zero-inflation, and batch effects). In this study, we analyzed orphan non-coding RNAs (oncRNAs) from serum samples of 1,050 individuals diagnosed with non-small cell lung cancer (NSCLC) at various stages, as well as sex-, age-, and BMI-matched controls to evaluate the potential use of deep generative models. We demonstrated that our multi-task generative AI model, Orion, surpassed commonly used methods in both overall performance and generalizability to held-out datasets. Orion achieved an overall sensitivity of 92% (95% CI: 85%–97%) at 90% specificity for cancer detection across all stages, outperforming the sensitivity of other methods such as support vector machine (SVM) classifier, ElasticNet, or XGBoost on held-out validation datasets by more than ∼30%.

Complete, telomere-to-telomere (T2T) genome assemblies promise improved analyses and the discovery of new variants, but many essential genomic resources remain associated with older reference genomes. Thus, there is a need to translate genomic features and read alignments between references. Here we describe a method called levioSAM2 that performs fast and accurate lift-over between assemblies using a whole-genome map. In addition to enabling the use of several references, we demonstrate that aligning reads to a high-quality reference (for example, T2T-CHM13) and lifting to an older reference (for example, Genome reference Consortium (GRC)h38) improves the accuracy of the resulting variant calls on the old reference. By leveraging the quality improvements of T2T-CHM13, levioSAM2 reduces small and structural variant calling errors compared with GRC-based mapping using real short- and long-read datasets. Performance is especially improved for a set of complex medically relevant genes, where the GRC references are lower quality.

Many bioinformatics methods seek to reduce reference bias, but no methods exist to comprehensively measure it. Biastools analyzes and categorizes instances of reference bias. It works in various scenarios, i.e. (a) when the donor’s variants are known and reads are simulated, (b) when donor variants are known and reads are real, and (c) when variants are unknown and reads are real. Using biastools , we observe that more inclusive graph genomes result in fewer biased sites. We find that end-to-end alignment reduces bias at indels relative to local aligners. Finally, we use biastools to characterize how T2T references improve large-scale bias.